agi-benchmarking-and-progress-metrics

As of mid-2026, the definition of Artificial General Intelligence (AGI) has transitioned from static, knowledge-based performance to agentic and interactive metrics Verified Answer #1. While the research community lacks a singular consensus on the exact definition of AGI, progress is measured through a suite of benchmarks testing expert knowledge, fluid intelligence, and autonomous execution Verified Answer #2. Legacy benchmarks such as MMLU are considered saturated, as Large Language Models (LLMs) routinely achieve over 90% accuracy on these historical metrics Verified Answer #2. Current industry tracking focuses on three primary pillars: fluid adaptability, expert-level knowledge synthesis, and sustained autonomous reliability Verified Answer #1.

Fluid Intelligence and Adaptation

Fluid intelligence is defined as the capacity to efficiently learn and adapt to novel tasks without relying on prior training data Verified Answer #2. The Abstraction and Reasoning Corpus (ARC-AGI) is the primary framework used to assess this trait Verified Answer #2.

• The ARC-AGI-3 benchmark, released in March 2026, requires agents to interact with novel environments, infer goals, and adapt in real-time Verified Answer #1.
• Humans typically score approximately 100% on these environments Verified Answer #1.
• Upon release, leading frontier models scored less than 1% on ARC-AGI-3, indicating a significant deficit in independent, goal-oriented exploration Verified Answer #1.
• On the earlier static ARC-AGI-1, OpenAI’s GPT-5.5 Pro achieved 96.5% by utilizing test-time compute scaling Verified Answer #2.

Expert Academic Synthesis

To evaluate the boundaries of specialized knowledge, researchers utilize Humanity's Last Exam (HLE) Verified Answer #1. This benchmark consists of 2,500 highly difficult, expert-vetted academic questions designed to resist simple internet retrieval Verified Answer #2.

• Human experts maintain approximately 90% accuracy across the graduate-level disciplines tested in HLE Verified Answer #1.
• As of March 2026, frontier systems consistently cluster between 40% and 50% accuracy Verified Answer #1.
• Google's Gemini 3.5 Pro achieved a score of 45.9%, while xAI's Grok 4 reached 44.4% Verified Answer #2.
• This performance gap suggests that current models struggle with cross-disciplinary reasoning that exceeds basic pattern matching Verified Answer #1.

Long-Horizon Autonomous Agency

Modern metrics for AGI have shifted toward measuring agentic reliability over time rather than simple percentage scores Verified Answer #1. The METR Time Horizon is used to track the length of complex tasks an AI can complete without failure Verified Answer #1. This metric specifically measures the duration of human-expert hours an agent can sustain with 50% reliability Verified Answer #1. Researchers use these time-based evaluations to determine the practical limits of autonomous execution and agentic tool use Verified Answer #2.